In recent years, with the increasing demand for diagnosing autism spectrum disorder (ASD), automated detection methods based on electroencephalogram (EEG) signals have gained significant attention. However, existing approaches still face challenges in terms of accuracy, generalization ability, robustness, and interpretability. An improved ASD detection model, attention-CosCNN-LSTM-net (ACLNet), is proposed. A multi-dimensional attention mechanism is leveraged to enhance the focus on critical signals. A cosine convolutional neural network is combined to capture frequency features of EEG signals, and a tree-structured LSTM module is integrated to model the hierarchical structure and long-term dependencies within signals. These components enable comprehensive extraction of spatiotemporal and frequency domain features of EEG signals. Experiments conducted on an ASD EEG dataset with five-fold cross-validation demonstrate that ACLNet achieves a classification accuracy of 94.11%, a recall rate of 93.29%, and a precision rate of 93.78%, significantly outperforming existing detection methods. Moreover, the model exhibits stable performance across different data splits and unseen data, validating strong generalization ability and robustness. Ablation studies further confirm the critical contributions of each module to feature extraction and overall performance. This study provides an efficient, stable, and interpretable solution for automated ASD detection, advancing the application of EEG signals in ASD diagnosis and offering valuable support for related research and clinical practice.